Neural State Classification (NSC) [19] is a scalable method for the analysis of hybrid systems, which consists in learning a neural network-based classifier able to detect whether or not an unsafe state can be reached from a certain configuration of a hybrid system. NSC has very high accuracy, yet it is prone to prediction errors that can affect system safety. To overcome this limitation, we present a method, based on the theory of conformal prediction, that complements NSC predictions with statistically sound estimates of prediction uncertainty. This results in a principled criterion to reject potentially erroneous predictions a priori, i.e., without knowing the true reachability values. Our approach is highly efficient (with runtimes in the order of milliseconds) and effective, managing in our experiments to successfully reject almost all the wrong NSC predictions.
